feat: Add logprobs support to SGLang backend

components/src/dynamo/sglang/request_handlers/llm/decode_handler.py

@@ -14,6 +14,8 @@
 from dynamo.sglang.publisher import DynamoSglangPublisher
 from dynamo.sglang.request_handlers.handler_base import BaseWorkerHandler
 
+logger = logging.getLogger(__name__)
+
 
 class DecodeWorkerHandler(BaseWorkerHandler):
     """Handler for decode workers in both aggregated and disaggregated serving modes."""
@@ -77,6 +79,7 @@ def _build_sampling_params(self, request: Dict[str, Any]) -> Dict[str, Any]:
             # Token-based request format
             sampling_opts = request.get("sampling_options", {})
             stop_conditions = request.get("stop_conditions", {})
+            output_options = request.get("output_options", {})
 
             param_mapping = {
                 "temperature": sampling_opts.get("temperature"),
@@ -85,6 +88,23 @@ def _build_sampling_params(self, request: Dict[str, Any]) -> Dict[str, Any]:
                 "max_new_tokens": stop_conditions.get("max_tokens"),
                 "ignore_eos": stop_conditions.get("ignore_eos"),
             }
+
+            # Handle logprobs from output_options
+            logprobs_value = output_options.get("logprobs")
+            if logprobs_value is not None and logprobs_value != "":
+                try:
+                    parsed_logprobs = int(logprobs_value)
+                    if parsed_logprobs < 0:
+                        logger.warning(
+                            f"Invalid logprobs value: {logprobs_value} (must be non-negative), ignoring"
+                        )
+                    else:
+                        param_mapping["return_logprob"] = True
+                        param_mapping["top_logprobs_num"] = parsed_logprobs
+                except (ValueError, TypeError):
+                    logger.warning(
+                        f"Invalid logprobs value: {logprobs_value} (must be integer), ignoring"
+                    )
         else:
             # OpenAI request format
             param_mapping = {
@@ -94,6 +114,14 @@ def _build_sampling_params(self, request: Dict[str, Any]) -> Dict[str, Any]:
                 "max_new_tokens": request.get("max_tokens"),
             }
 
+            # Handle logprobs from OpenAI format
+            logprobs = request.get("logprobs")
+            top_logprobs = request.get("top_logprobs")
+            if logprobs:
+                param_mapping["return_logprob"] = True
+                if top_logprobs is not None:
+                    param_mapping["top_logprobs_num"] = top_logprobs
+
         return {k: v for k, v in param_mapping.items() if v is not None}
 
     async def generate(
@@ -193,6 +221,94 @@ async def generate(
                 async for out in self._process_text_stream(agg, context):
                     yield out
 
+    @staticmethod
+    def _extract_logprobs(
+        res: Dict[str, Any], num_output_tokens_so_far: int
+    ) -> tuple[list[float] | None, list[list[dict]] | None]:
+        """
+        Extract logprobs from SGLang response for new tokens.
+
+        Args:
+            res: SGLang response dict
+            num_output_tokens_so_far: Number of tokens already processed
+
+        Returns:
+            Tuple of (log_probs, top_logprobs) in Dynamo's expected format:
+            - log_probs: List of log probabilities for each new token
+            - top_logprobs: List of top logprobs dicts for each new token
+        """
+        meta_info = res.get("meta_info", {})
+
+        # SGLang uses "output_token_logprobs" for selected token logprobs
+        # Format: [(logprob, token_id, decoded_text), ...] - one tuple per token
+        output_token_logprobs = meta_info.get("output_token_logprobs")
+
+        # SGLang uses "output_top_logprobs" for top-k alternatives
+        # Format: [[(logprob, token_id, text), ...], ...] - list of lists
+        output_top_logprobs = meta_info.get("output_top_logprobs")
+
+        if not output_token_logprobs:
+            return None, None
+
+        # Get logprobs for new tokens only
+        new_token_logprobs = output_token_logprobs[num_output_tokens_so_far:]
+        if not new_token_logprobs:
+            return None, None
+
+        log_probs = []
+        top_logprobs = []
+
+        # Get top logprobs slice if available
+        new_top_logprobs = None
+        if output_top_logprobs:
+            new_top_logprobs = output_top_logprobs[num_output_tokens_so_far:]
+
+        # Extract logprobs for each token, maintaining 1:1 alignment
+        for idx, token_data in enumerate(new_token_logprobs):
+            # Skip if token_data is None or logprob_val is None
+            if token_data is None:
+                continue
+            # SGLang format: (logprob, token_id, decoded_text)
+            logprob_val = token_data[0]
+            if logprob_val is None:
+                continue
+
+            log_probs.append(float(logprob_val))
+
+            # Extract corresponding top logprobs for this token position
+            if new_top_logprobs and idx < len(new_top_logprobs):
+                token_top_list = new_top_logprobs[idx]
+                if not token_top_list:
+                    top_logprobs.append([])
+                else:
+                    # Filter out None entries and sort by logprob descending
+                    # SGLang doesn't guarantee order, so we sort to assign proper ranks
+                    valid_entries = [
+                        alt_data
+                        for alt_data in token_top_list
+                        if alt_data is not None and alt_data[0] is not None
+                    ]
+                    # Sort by logprob descending (highest probability first)
+                    valid_entries.sort(key=lambda x: x[0], reverse=True)
+
+                    token_top_logprobs = []
+                    for rank, alt_data in enumerate(valid_entries):
+                        # SGLang format: (logprob, token_id, decoded_text)
+                        alt_logprob_val = alt_data[0]
+                        token_id = alt_data[1]
+                        decoded_text = alt_data[2] if len(alt_data) > 2 else None
+                        token_top_logprobs.append(
+                            {
+                                "rank": rank,
+                                "token_id": token_id,
+                                "token": decoded_text,
+                                "logprob": float(alt_logprob_val),
+                            }
+                        )
+                    top_logprobs.append(token_top_logprobs)
+
+        return log_probs if log_probs else None, top_logprobs if top_logprobs else None
+
     async def _process_token_stream(
         self,
         stream_source: AsyncGenerator[Dict[str, Any], None],
@@ -239,6 +355,16 @@ async def _process_token_stream(
 
                 next_total_toks = len(output_ids)
                 out["token_ids"] = output_ids[num_output_tokens_so_far:]
+
+                # Extract logprobs for new tokens
+                log_probs, top_logprobs = self._extract_logprobs(
+                    res, num_output_tokens_so_far
+                )
+                if log_probs is not None:
+                    out["log_probs"] = log_probs
+                if top_logprobs is not None:
+                    out["top_logprobs"] = top_logprobs
+
                 num_output_tokens_so_far = next_total_toks
                 if finish_reason:
                     input_tokens = res["meta_info"]["prompt_tokens"]

tests/serve/test_sglang.py

@@ -17,7 +17,9 @@
 from tests.utils.payload_builder import (
     chat_payload,
     chat_payload_default,
+    chat_payload_with_logprobs,
     completion_payload_default,
+    completion_payload_with_logprobs,
     embedding_payload,
     embedding_payload_default,
     metric_payload_default,
@@ -229,6 +231,33 @@ class SGLangConfig(EngineConfig):
             completion_payload_default(),
         ],
     ),
+    "aggregated_logprobs": SGLangConfig(
+        name="aggregated_logprobs",
+        directory=sglang_dir,
+        script_name="agg.sh",
+        marks=[
+            pytest.mark.gpu_1,
+            pytest.mark.pre_merge,
+            pytest.mark.timeout(240),  # 3x measured time + download time
+        ],
+        model="Qwen/Qwen3-0.6B",
+        request_payloads=[
+            chat_payload_with_logprobs(
+                repeat_count=2,
+                expected_response=["AI", "knock", "joke"],
+                max_tokens=30,
+                temperature=0.0,
+                top_logprobs=3,
+            ),
+            completion_payload_with_logprobs(
+                repeat_count=2,
+                expected_response=["AI", "knock", "joke"],
+                max_tokens=30,
+                temperature=0.0,
+                logprobs=5,
+            ),
+        ],
+    ),
 }